Everything built on Jev, the System One model from TypeSafe AI — the first model that answers with typed judgments and calibrated probabilities instead of prose. 352 projects, three languages, refreshed from the GitHub API.
billing probabilities {"billing":0.94,"technical":0.05,"other":0.01} One request, three conditions, three typed answers. The page you are reading is the table. curl -X POST https://api.typesafe.ai/v1/systemone \
-H "Authorization: Bearer $JEV_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"state": { "document": "I was charged twice. Please fix this now." },
"questions": {
"category": {
"type": "choice",
"question": "What is this ticket about?",
"options": ["billing", "technical", "other"]
}
}
}'A decision table gets cited by rule number. So does this one.
Every entry below is generated from one data file. Stars, licenses and push dates refresh from the GitHub API on a schedule, and every row is link-checked before it ships. Star counts are a dated snapshot, not a ranking claim.
The spec sheet, the three primitives and the docs — read these before building.
Maintained by TypeSafe AI, plus the frameworks that ship Jev natively.
Routers, tool gates, reviewers, skills and MCP servers for agent harnesses.
Per-turn model and tool selection, under a hard deadline.
Decide what stays in the window before the model ever reads it.
Judges, linters, coverage gates and review dashboards.
Jev picks the operation and the DOM element; a small LLM only writes the text.
Driving phones, IM clients and native UIs without hooking or patching.
Query understanding, source selection, reranking and semantic SQL.
Guardrails, prompt-injection checks, judges that abstain.
Postgres extensions, semantic SQL and telemetry pipelines that call Jev.
End-user tools people actually open every day.
Unofficial clients for the languages without a first-party SDK.
Call Jev from a shell, no SDK required.
Measure it before you trust it.
Run System One semantics without the vendor — on a 3090 if you like.
Decisions as game mechanics.
Scoring financial signals with calibrated probabilities.
Try it in thirty seconds.
It is not a hand-typed list that rots in six months. It is one data file, deterministic scripts, and a free-AI reviewer. No API keys: the maintainer runs on free open-source models.
Two-model gate: a second, independent free model reviews every automated pull request, and a change that touches only the report is never merged. If no model is reachable, the scripts degrade to deterministic reports instead of failing.
A System One model from TypeSafe AI. You send a state plus typed questions, and it returns typed answers with a probability per option and a confidence value. Model jev-latest, one POST to /v1/systemone, 70–500 ms end to end.
An LLM returns text you have to parse, and every parse is a place to fail. Jev returns a typed choice plus a probability distribution, so you can set a threshold, unit-test the branch, and log a number.
Only for the hosted API. Rule 229 collects the open-weight and replica implementations that run System One semantics without the vendor.
Open a pull request against data/entries.json, or tag your repository jev and open an issue. One entry per repository, one PR per entry, and it must ship a description in all three languages.
No. It is community-maintained. TypeSafe AI is linked because Jev is their model, and the distinction is stated on the page rather than buried in a footer.
Yes. Everything here is CC0-1.0 public domain. Fetch projects.json for structured records, llms-full.txt for every entry as markdown, or one category at a time from c/<rule>.md.